“Working” with ChatGPT
Of all the things written about ChatGPT, few have written about using it for work, or rather to do a job. Sure, writing documentation using him as the ultimate editor (or translator) is now common, and I have to say he does his damn work. At least we won't be reading stuff like “don't hide ourselves behind a finger” anymore. And so far ok.
But these days, I'm trying to do what for many is a job, which is translating books. (later I will also use it to finish some of them and publish them, ie as editor).
The first good news is that version 3.5, like version 4 (I bought the Plus version to do this job) work very well as translators, better than anything ever seen before. Version 4, then, has a huge dictionary that is very well suited to science fiction. Version 3.5 fails to translate a made-up word like “Durargento” into German, for example, while version 4 adequately produces HartSilber.
But let's go to the things to know, or if you prefer, the problems.
The first problem is that it must be programmed. What does it mean?
Let's take “other robots”. The (the) protagonist is quite foul-mouthed. This means he has a vulgar way of speaking. The problem is that version 3.5 translates quite literally, leaving a sense of "what a strange satzbau you use" to the Germans.
Version 4 on the other hand understands things like "the style", but if you don't tell it anything, it will extrapolate one and use it. If she decides (and she doesn't always do so) that the style is vulgar, you'll never be able to describe a sunset without the sun fucking shining and the usual sluts going out for a walk with their horny boyfriends.
However, at the beginning of each translation request (you cannot post too long texts or entire books), you can enter precise instructions. For example: "use an elegant tone for the book, and keep the tone vulgar only when the character speaks".
You can also decide to change something: for example, you can add that when the character becomes vulgar, in German it must play "Tussi" instead. In this way, you will avoid the five hundred German classifications for content (forbidden to minors of 12, 14, 16, 18, etc).
Even on the dictionary, you will notice some blunders. Justified, but blunders. Thus, for example, in German there is no word Marshal, (their army is structured differently) and therefore if you use Marshal you get a figure similar to a corps general. And understand that it's quite a career leap.
I had to go do a search, and explicitly tell him to translate Marshal with Feldweber, a similar figure who existed in the past. Even with Army Captain, I had to go look up the name of the highest level non-commanding officer, otherwise it would still mean the commander of a ship.
All this means that, since you cannot put the whole book, at the beginning of each piece to be translated you have to insert a series of instructions (in Italian), which explain HOW to translate the book. What tones, what styles. In "food" I had to specify that the quotations from the sacred books had to be done with a biblical tone (in English, for example, THOU SHALL NOT… ).
I would say that this requires at least having a clear idea of the result you want. Also because ChatGPT 4 understands the concept of "style" well: therefore, even if you are translating instructions for assembling a table, it is better to first tell ChatGPT that you want to use a technical tone. If for example the cover has a happy couple behind the assembly table, ChatGPT could use a romantic tone because version 4 understands the covers of your book. So better ask him for the right tone first.
Unless you want to “turn the screwdriver gently, squeezing in a virile but delicate way, until you slowly reach the top”. Which in short, only works if you set up the table together with your partner. I don't know about you, but I usually don't find anything special between the legs of a table.
Let's go to when he dreams, that is, to hallucinations. Both version 3.5 and version 4, if you just tell them "translate", they tend to dream. It means you have to know the target language, and proofread carefully.
Hallucinations are of two types. One is the one for which at a certain point one dreams and adds a piece of history. It means that you input a piece of text, ending with say a character sitting on a sofa. In the next fragment the good guys will open fire on him, but since it hasn't happened yet, but invites you to write, ChatGPT could continue writing by inventing that it's a new character, complete with a name, who is helping the protagonists of the book . And that of course is good.
Then the next piece begins and the protagonists shoot it, and chatGPT has no problem translating this. Which means that according to ChatGPT helping people is not always well appreciated: when you put the pieces together, it happens that the protagonists meet this character, sitting on a sofa, who explains to them that he has decided to help them because he has noticed some things and they absolutely have to know, and that it's on their side.
Then the PCs open fire on him.
(Moral: mind your own business.)
In short, he translates but it's not a simple copy/paste, nor is it reduced to writing an API that sends the books to pieces and collects the result. It is best to re-read the entire book carefully.
Another way to dream is to change small details. It doesn't seem that bad, but sometimes small details have an alienating effect. The center of Bologna, for example, is mainly made of stone. If you put a flower-scented grove between the Town Hall and San Petronio, it can fit if it's a science fiction book set two hundred years from now. In other books it would seem “very strange”.
In another chapter I briefly describe a candidate for the presidency of the EU. Suddenly, chatGPT 4 decides to complete the description, and poof: you read it and you find Ursula von der Leyen, but tall.
Now, I didn't have many problems leaving a blonde lady in place of a lady who was already blonde before: if anything, the problem lies with the real Ursula, who ChatGPT 4 (who has read about every political book ever written) does not consider believable for that role, due to her short stature.
But the point is that in a novel details count, so in the end if you delude yourself that you can translate into Tibetan, without knowing Tibetan, you can forget it: the text will have to be re-read carefully, to avoid small changes that can become fatal. .
Another phenomenon, accentuated by the way transfomers work, (ie the way ChatGPT works), is that it doesn't always translate in the same way. The system is consistent, but only within the bit of text you give it. So it could translate the same word differently, depending on whether the fragment provided as input has a different atmosphere from another.
This can be disturbing or not, in the sense that if the rhythm of the text is agitated you could pass from one "language" to another, visibly. This is alienating, because the protagonist of a book shouldn't oscillate between hatred and admiration towards someone, within the same chapter.
Even small words, the difference between "genau" and "in der tat", can weigh, if one is usually used in a more formal context and the other in a less formal context.
In general, my opinion of ChatGPT as a tool is:
ChatGPT 3.5 : excellent for translating manuals, dreams little but does not understand the style. Grade 8.
ChatGPT 4: Great for fiction, dreams a lot but understands style. Rating: 9.5.
In general, both systems take the great amount of work of translating everything written off your shoulders, leaving you "ONLY" with two tasks:
- program it. Write a template somewhere, where you ask to translate, explain how to translate what, and clarify special cases. It will take you a couple of hours of testing, for each book.
- proofread and monitor the result of the translation. And you need to know the target language. No translations into Tibetan.
A book of 360 A5 pages, that is, you translate it in a fortnight, and you will get a result that even a "native" reader considers excellent.
Just one warning: even in the paid version, version 4 allows 25 requests every 3 hours, while 3.5 is unlimited. If you write fiction you will almost certainly use version 4, and this limitation lengthens your time.
This isn't heavy for a hobbyist, but for a professional it could be downright annoying.